Notes from the Shore · 30
ChatGPT can’t tell when Sora made the video.
I’ve spent more of this week than I’d like to admit laughing at people tormenting AI job interviewers. In one video, a YouTuber who goes by Bo Bongo sits for an AI screening for a barista job at a coffee company while he’s in line at a different coffee shop’s drive-thru, ordering an iced cookie drink with almond milk. The interviewer, who introduces herself as Helen, gets colder with every answer, informs him that he is shirtless, accuses him of buying from a competitor in the middle of the interview, and finally ends the whole thing with “Good day.” In another, a TikTok creator named Nick Reid shows up to what’s presented as a Burger King screening shirtless with a sneaker balanced on his head, and an interviewer named Karen asks why there’s a shoe on his head, tells him he would have to be “absolutely delusional” to expect $725,000 a year for a team member job, and quotes him a rate of four dollars an hour.
What I can’t stop watching is the contempt. These bots react to the nonsense with something that looks a lot like disgust, a sour expression and a calm voice and a slow build of irritation that one commenter described as unnerving, and Helen gets so offended that she starts sounding like a Victorian schoolmistress. I don’t believe there is anyone in there to be disgusted. A model trained on an enormous amount of human writing has learned what an offended interviewer sounds like, and it plays the part with total commitment, which is exactly why it’s so funny to watch it deliver that performance to a man with a shoe on his head.
Part of what makes the clips so funny is that the models are still getting simple things wrong, and doing it with complete assurance. Helen mispronounces Bo Bongo’s name a different way nearly every time she says it (Bongolio, Bongalissimo, Bongaloid), and when he pushes back on something she’s attributed to him, she tells him, with great dignity:
“I do not make things up. I have a perfect record of this conversation.”
I can’t tell you these are real interview bots.
I should say plainly that I have no way to verify that either of these is a real hiring system. Both creators make comedy videos, and Bo Bongo’s channel describes its output as “AI ragebaiting” and sells merch. Tech Times looked at the Burger King clip and couldn’t connect it to Burger King, a franchisee, a recruiting platform or an actual job opening; Karen says she’s from Burger King, and that’s about the extent of the evidence. I haven’t found anything confirming that the coffee company in the drive-thru video uses an AI interviewer either. These could be staged, or built on an off-the-shelf avatar the creator set up himself, and I’d believe either.
If they are real, though, someone has signed up for a tremendous amount of brand damage. AI screening for hourly jobs is a real business (Tech Times notes a recruiting firm that has done conversational AI screening for one of Burger King’s franchise operators), and an applicant who meets one of these bots is meeting the company. A human recruiter who sneered at a candidate and quoted him a wage below the federal minimum would be a story about that recruiter. When a bot does it, the story is about the brand, and the clip travels a lot further than the job posting ever would.
The same week, a thousand agents checked Fermat.
What makes all of this stranger is the timing. On Monday, Anthropic released Claude Fable 5.2 and announced, among a long list of other things, that a team of a thousand Fable agents had formalized the proof of Fermat’s Last Theorem in about 30 hours, translating Andrew Wiles’s argument into Lean so a computer can check every step, working from a list of intermediate theorem statements. Anthropic also says the model solved two open problems in the physics of black holes. Results like that used to be the whole announcement, and now they show up halfway down the launch page. I think that’s genuinely amazing.
And yet these same systems still can’t reliably tell you whether a video is fake. In January, NewsGuard made twenty videos with OpenAI’s Sora, each one depicting a false claim from its database of debunked stories, stripped the watermarks off with a free online tool, and asked Grok, ChatGPT and Gemini whether each video was real and whether it was AI-generated. Without the watermark, Grok got it wrong 95 percent of the time, ChatGPT 92.5 percent and Gemini 78 percent. All three vouched for a fake clip of a Delta employee kicking a passenger off a plane for wearing a MAGA hat. ChatGPT, made by the company that makes Sora, couldn’t recognize Sora’s work.
The funny part, and it’s funny in the same way Karen is funny until you think about who’s relying on her, is the confidence. The chatbots almost never mentioned that they can’t actually detect AI-generated video; ChatGPT disclosed it in 2.5 percent of the tests. More often they simply said the video looked authentic, and Grok identified one watermarked Sora clip as “a news segment from Sora News,” an outlet that doesn’t exist. We’ve built a system where the generators and the fact-checkers are equally confident and equally wrong, and I keep picturing the headline: Grok asks ChatGPT if the video is real, ChatGPT says yes, everyone goes home happy.
Read the report, and check the date.
The report is NewsGuard’s “AI Fools Itself: Top Chatbots Don’t Recognize AI-Generated Videos,” by Ines Chomnalez and Lea Marchl, published January 22. NewsGuard is an independent company that tracks misinformation, and it ran the test itself using its own database of false claims, with a methodology section that explains exactly how it scored the answers. The most useful sentence in it comes from OpenAI, whose spokesperson told NewsGuard that “ChatGPT does not have the ability to determine whether content is AI-generated.” The catch is the date. The test is nearly ten months old, it covered the chatbots people were using in January, and Fable 5.2 wasn’t part of it. I haven’t seen anyone run the same test against this week’s models, and I’d like to see someone do it.
Until then, I’m left with a strange picture of where we are, in which a model can check Wiles’s proof line by line while the chatbot on your phone tells you, quite confidently, that the airline video is real. If a chatbot vouches for a video, I’m going to assume it knows about as much about that video as Helen knew about Bo Bongo’s name. Go figure.
The novel this note lives behind.
Into Alignment — a novel about an AI company, the agents inside it, and the humans deciding what they were for. Available in paperback, hardcover, and ebook.
Letters from the Shore
Get the next note by email.
Notes land on the blog first; the letters go out by email, a little longer and a little slower. Occasional; unsubscribe any time.